Skip to content

Why ParleML

The enterprise-grade ML platform, built to be driven by AI agents. Free.

The modelling the commercial ML platforms charge for, free: deep feature engineering, stacked ensembles, forecasting, anomaly detection, regulated-model templates and interpretation down to every rule. Your own AI agent does the work and you make the decisions.

docker run -d --name parleml -p 8000:8000 -v parleml-data:/app/data_root ghcr.io/parleml/parleml:0.8.0

Free for commercial use · agents on your subscription, an API key or a local model · no telemetry · runs on your own machine or servers, even fully offline · your data stays there

  • The modelling enterprise platforms charge for

    Deep feature engineering, strong free-text handling, stacked ensembles or deep hyperparameter search, foundation models, regulated-model templates, reason codes, fairness, AutoDoc model cards and fast Treelite scoring. All built in, not an add-on.

  • And interpretation that goes further

    The full standard set (SHAP, partial dependence, ICE, ALE, surrogates, what-if), then archetypes, rules tested on unseen rows, interactions and trust checks. See the groups the model sees, which rules hold up on rows the model never saw.

  • Eight kinds of ML problem, one platform

    Classification, regression, forecasting, ranking, recommendation, basket rules, clustering and anomaly detection. Every one of them explained.

  • Your agent becomes a senior data scientist

    Claude Code, Codex, Gemini CLI or Cursor don't write throwaway ML code: they drive a full ML platform. They see every result and explanation, work through tested tools and skills, and hand the decisions to you. Run them inside ParleML with your ChatGPT, Claude or Google AI plan, an API key or a local model, or plug ParleML into the agent you already use.

Get started in 5 minutes Train your first model

The long version

Open any part for the detail.

One engine, every modelling strategy

ParleML builds on AutoGluon, LightGBM, XGBoost, CatBoost and neural networks rather than writing yet another booster. AutoGluon's multi-layer stacking is one option, not the only one. Each experiment uses a template, and every template can be opened and changed:

You want Templates
A sensible start Default, Quick Explore
Top accuracy by breadth and stacking High Accuracy, Diverse Ensemble
Deep tuning, the Driverless AI and mljar approach: an in-depth search on one or several model families, with an ensemble of the best on top Tuned Single, Tuned Ensemble, Reliable Tuned Ensemble, Tuned Variants
Foundation models Foundation Ensemble (TabICL as an ensemble member)
A model a regulator can read Regulated Glassbox (GLM, EBM or RuleFit), Regulated Scorecard (weight of evidence, coefficients as points), Regulated Monotonic
Fast scoring Real Time: one tuned booster, scored through Treelite
Your data's shape Large Dataset, Small Data, Text Heavy

Underneath: grouped and time-aware validation, a sealed test set, Optuna search, class-imbalance treatment chosen by search (class weights), SMOTE-family resampling, automatic monotonic constraints, sample weights, GPU training and calibration (isotonic, Platt, temperature).

Feature engineering: the standard range, then our own

Driverless AI's real strength is its range of automatic transformers. Going by its 2.5.1 documentation (as of 18 October 2026), ParleML has a counterpart for nearly every one, then goes further:

Driverless AI transformers ParleML
CVTargetEncode, NumCatTE, NumToCatTE, WeightOfEvidence, NumToCatWoE (monotonic) Six target encoders (mean, weight of evidence, James-Stein, M-estimate and others) and a count encoder, fitted inside the folds
Interactions, InteractionsSimple Tree-guided interactions: the pairs and operations chosen by a model, with an optional feedback round
CVCatNumEncode Group aggregates of a number within a category, plus within-group percentile ranks
ClusterId, ClusterDist, Binner Clustering encoder and learned bins
OneHotEncoding, Frequent, LexiLabelEncoder and SortedLE, StandardScaler, QuantileTransformerNormal and QuantileTransformerUniform Category handling, contrast coding (Helmert, backward difference, sum), power transforms, scaling for linear models
Dates, IsHoliday, DatePolar Calendar parts and cyclical encodings, plus optional holiday and business-hours flags
Lags, EwmaLags, LagsAggregates Per-entity lags, EWMA, rolling windows and time since the last event, when you switch them on
Text, TextLinModel, BERT, TextCNNV2 and TextBiGRUV2, TextClustTE and TextClustDist See the text section below

Beyond what we found in the other platforms:

  • Features from domain columns: email, IP location (with a MaxMind database you add) and network, URL, user agent, device info, phone, card number, IBAN, VAT, street address, postal-code distances, personal names, and cross-field consistency checks. ParleML or your agent suggests each column's type; once it is accepted, these features switch on.
  • Messy categories: MinHash and Gamma-Poisson encoders for typos and very high cardinality.
  • Tag lists and packed counts: see the text section below.
  • Features from your database: a knowledge base finds how your tables join (you confirm the inferred joins), and every SQL recipe, yours or your agent's, is checked so no feature reads data from after the prediction moment.
  • Entity fan-out: how many distinct values an entity links to (cards per device, for example).
  • Honest selection: a "no better than noise" reference that flags weak features, and feature pruning that is undone when it costs accuracy.

Driverless AI also offers image features and ELM features, which ParleML doesn't have yet. Isolation forest, which Driverless AI uses to add an anomaly-score column, is in ParleML a full anomaly detector, one of 23.

Free text, handled properly

Most platforms treat a text column as prose. ParleML tells four kinds of text column apart and gives each the encoder that fits it:

Kind of column Example What ParleML does
Prose complaint notes, claim descriptions TF-IDF topics plus a cross-validated linear text signal, or neural sentence embeddings
Tag and keyword lists action|comedy|drama Every token target-encoded on its own and combined per row, per class for multiclass, fitted inside each fold
Packed counts Netflix(2), McAfee(1) Tokens and their counts parsed and compressed into a few topic columns
Messy short strings job titles, merchant names with typos Gamma-Poisson topics or MinHash fingerprints of character patterns

We haven't found encoders for tag lists or packed counts in the other platforms we compared: a standard encoder sees each such cell as one rare category and learns almost nothing from it.

Around the encoders:

  • Embeddings your way. A model on your server (English or multilingual, nothing leaves the machine) or your own embeddings provider. The fit records the exact model, and scoring checks the provider still embeds the same way, so a silent model swap can't change your predictions.
  • Text Heavy template for data whose signal is in free text.
  • Text in clustering and anomaly detection, fused with the other columns, each cluster described by its keywords.
  • Before training: a word cloud for any text column.
  • After training: for every text encoder, the words behind each component and each outcome, and the words that separate each outcome from the rest.

As documented, Driverless AI covers prose with TF-IDF, a linear text model, CNN and GRU networks and BERT. It also explains single rows token by token (NLP LOCO), which ParleML doesn't do yet.

Interpretation: everything standard, then our own

The standard views the commercial platforms ship are included (some also offer K-LIME or XEMP, which ParleML doesn't):

Question Views
What drives the model? Native, SHAP and permutation importance
How does each feature move the prediction? Partial dependence, ICE, ALE
Why this prediction? SHAP reason codes, LOCO, what-if and sensitivity
Can a simple model mimic it? Surrogate decision tree, residual surrogate
Is it fair? Disparate impact by any column you mark
Can I document it? AutoDoc model card

Then views we haven't found in the other platforms we compared:

View What it shows
Archetypes The groups inside each class the model sees, named, with members, stability and errors
Three rule engines Plain if-then rules, each with its Train and Test result; one engine maximises business value
SHAP interactions and compacity Which features act together, and how few features explain a decision
Archetype counterfactuals What would move a case from its risky group towards a safer one, and how much each feature changes
Trust summary Nine checks, from lift and calibration to drift, fairness and rule survival
Label-noise review Rows a model that never trained on them confidently disputes (confident learning, as in cleanlab), listed as candidates to review

Also included: an error tree that finds the groups of rows where the model is wrong most.

Forecasts, clusters, anomalies, ranking, recommendations and basket rules are explained too, each in the way that fits it. Classifiers and regressors trained elsewhere (scikit-learn, LightGBM, XGBoost, CatBoost) can be brought in and explained the same way.

Ready for regulated models
  • AutoDoc model card with data quality, features, performance, interpretability, risk and governance sections, and a crosswalk to the regulatory requirements it answers.
  • Glass-box templates: scorecard, GLM, EBM, RuleFit and monotonic boosters.
  • Fairness by any column you mark, without using it as a predictor.
  • Reason codes for each scored row, when you ask for them: its drivers and the rules it matched.
  • Evidence on held-out rows for every rule and diagnosis, so a reviewer can verify it.
  • Identifier pseudonymisation before any model sees card numbers, IBANs or emails.
  • Release records with sign-off, and an append-only log of changes, sign-ins and agent actions.

ParleML doesn't deploy or monitor models in production. Use it with your MLOps tooling: models export as Python packages and MLflow models.

Compared with SAS Viya, the incumbent in banks and insurers (from SAS's public pages, as of 18 October 2026): SAS leads on enterprise model risk management (model inventory and validation workflows), offered as its own product, with credit scorecards in a Risk Modeling add-on. ParleML covers much of the regulated modelling itself for free: weight-of-evidence scorecards, glass-box and monotonic models, fairness checks, reason codes, model cards and sign-off records.

Eight kinds of ML problem
Problem What you get
Classification and regression Binary, multiclass and regression, with calibration, thresholds and business value
Forecasting Four backends, 28 named models plus AutoGluon's zoo: 19 statistical (ARIMA, ETS, Theta, CES, TBATS, MSTL, intermittent demand), LightGBM, XGBoost and CatBoost on lags, six Chronos foundation checkpoints, and AutoGluon-TimeSeries. Race them per series, with prediction intervals
Ranking Learning to rank with LambdaRank, tuned for NDCG
Recommendation Next best product, with calibrated models per product and customer group
Basket rules Association rules tested on held-out baskets, corrected for multiple testing
Clustering Seven algorithms, the number of clusters chosen for you, clusters named
Anomaly detection 23 detectors with an automatic mode, and the reason each row was flagged
Honest by design
  • A model that can't beat a no-skill baseline is flagged, and the floor is shown before you train.
  • An ensemble ships only if it beats the simpler model.
  • Leaks are flagged before training, and a confirmed leak stays confirmed on derived data.
  • Database features are checked so that none reads data from after the prediction moment.
  • Every rule and diagnosis is reported on rows the model never trained on.
  • When something can't be done properly, ParleML says so in words instead of returning a number you shouldn't trust.
Agents: we don't sell an agent, we make yours better

Agents can already do a lot. What they lack is a reliable ML engine and the evidence to make good decisions. ParleML gives them both, in both directions:

Your choice of access. The agents inside ParleML are the real Codex, Claude Code and Gemini CLI. Sign them in with a ChatGPT, Claude or Google AI subscription, use an API key, or run a local model with Codex (Ollama or LM Studio) or Claude Code (Ollama). A tip: a long ML session reads a lot of results, so a flat monthly plan can cost far less than per-token billing.

Agents inside ParleML ParleML inside your agent
Codex and Gemini CLI in the standard image, Claude Code and Cursor in the extended image; local models with Codex (Ollama or LM Studio) or Claude Code (Ollama) An MCP server with 26 tools in three permission levels: read, compute, builder
Sessions that frame, build, challenge and release models, with a live flow map Interactive ParleML panels inside Claude Desktop, ChatGPT, VS Code and Cursor
Decisions go to people. Who answered, and when, is recorded Sign in from claude.ai or ChatGPT, or use an API token; build tokens are bound to chosen workspaces
17 shipped skills, plus your own skills, rules and hooks The same skills for your own agent, exported from Settings → Skill packs

Every object records the session and step that made it. Plans are reviewed before launch, can carry a stop rule, and a launch budget caps how many runs, and how large, an agent may start. You can work through the agent, the classic pages, or both.

Why not just ask a coding agent?

A capable coding agent with scikit-learn and LightGBM can train you a model in minutes. What it doesn't give you on its own:

A coding agent on its own The same agent with ParleML
Writes new ML code every session, and you review that code Calls tested tools and skills; you review results and decisions, not code
Explanations it improvises: a SHAP plot, if you ask Interpretation computed for every model, in every kind of problem
Checks it may or may not remember Leakage, the no-skill baseline, held-out tests for every rule and point-in-time checks, every time
Choices made silently in code Decisions handed to you, with who answered and when on record
Results scattered over notebooks and chat history Every dataset, experiment, model and decision kept in the workspace
Its context window fills with code and stack traces Its context holds results and evidence, so it reasons further

It's the same agent, with the same access. ParleML makes its work reliable, repeatable and reviewable.

How ParleML compares

About these comparisons

The comparisons below are our reading of each product's public documentation, release notes and product pages, as of 18 October 2026. They are offered in good faith, to the best of our knowledge, to help you choose; they are not a test of the products. "Not found in public docs" means we found no mention, not that a product can't do it. Products change often, and we may have missed or misread something. If anything here is wrong or out of date, write to [email protected] and we will correct it as soon as we can confirm it. Product names are trademarks of their owners; ParleML is not affiliated with or endorsed by any of them. The sources are listed at the end of this section.

How it compares with Driverless AI, DataRobot and Dataiku

As of 18 October 2026, from each vendor's documentation, release notes and product pages:

ParleML Driverless AI DataRobot Dataiku
AI agents Claude Code, Codex, Gemini CLI, Cursor or a local model run governed ML sessions on your own access; MCP with interactive panels No built-in agent; GPT summaries of datasets and experiments (off by default); H2O's separate h2oGPTe agents can run experiments once an administrator connects them Platform to host your company's agents; a data science agent, mainly for data preparation; open-source skills and an MCP server for coding agents Its own building agent (Cobuild), Claude Code and Codex in its Code Studios, agent hub
Price Free, including commercial use Commercial licence; demo or trial on request Commercial, no public price; 30-day trial Commercial, no public price; free on-premises edition; 14-day cloud trial
Interpretation The full standard set with ICE and ALE, plus archetypes, rules with Train and Test results, trust summary SHAP, K-LIME, PDP, ICE, surrogate tree, disparate impact; ALE only as a custom explainer recipe SHAP, XEMP, feature impact, partial dependence, fairness; ICE in a code accelerator; ALE not found in public docs SHAP and ICE-based row explanations, partial dependence, error tree, subpopulations, counterfactuals, fairness; ALE not found in public docs
Model documentation AutoDoc model card for every kind of experiment, written from the run's recorded evidence: up to 17 sections, from data quality and feature engineering to drift, limitations, monitoring plan and sign-off, with the workspace's intended use, and the model's risk level and approver from its release record. Mapped to the EU AI Act, SR 11-7, PRA SS1/23 and NIST AI RMF, with a list of the evidence still missing. PDF, Word, HTML or Markdown AutoDoc Compliance documentation with a template builder; states SR 11-7 support Model Document Generator (Word); an EU AI Act readiness solution
Data checks Leakage, no-skill floor and a signal check before training (TabICL; Chronos for forecasts); mislabelled-row review; column meanings such as email, phone, address and IP suggested for you to accept Leakage detection, on by default Target-leakage and data-quality checks Column meanings such as email, IP, URL and country; after training, leakage warnings and a check against a dummy model
Ranking Learning to rank Not found in public docs Not found in public docs Not found in public docs
Recommendation Calibrated next best product Not found in public docs Recommendation engine as a code accelerator (multilabel model) SQL collaborative-filtering plugin
Basket rules Built in, tested on held-out baskets Not found in public docs Not found in public docs Packaged industry solution
Forecasting Four backends, 28 named models plus AutoGluon's zoo: 19 statistical (AutoARIMA, AutoETS, AutoTheta, AutoCES, MFLES, TBATS, MSTL, Croston, ADIDA, IMAPA, TSB and baselines), LightGBM, XGBoost and CatBoost on lags, six Chronos-2 and Chronos-Bolt checkpoints, and AutoGluon-TimeSeries (DeepAR, TFT, Toto-2, tabular). Raced per series, with prediction intervals Lag features on its standard models ARIMA, ETS, TBATS, Prophet, LSTM, DeepAR, GBMs AutoARIMA, Prophet, DeepAR, Transformer, MQ-CNN, baselines; XGBoost, random forest and ridge on lags
Anomaly detection 23 detectors (ECOD, COPOD, isolation forest, LOF, one-class SVM, autoencoder, VAE and more), an automatic mode, and the reason each row was flagged Isolation Forest Isolation Forest, LOF, one-class SVM, MAD, Mahalanobis, autoencoders Isolation Forest
Feature engineering Counterparts for nearly all of Driverless AI's, plus domain columns, tag lists, multi-table recipes checked point in time 50+ transformers Multi-table feature discovery Pairwise interactions, encodings; genetic feature generation as a plugin
Free text Four kinds of text column, each with its own encoder; words behind each outcome TF-IDF, linear model, CNN, GRU, BERT; per-row token explanations TF-IDF, small transformers, CNN; text explanations TF-IDF, hashing, sentence or LLM embeddings
Glass-box and scorecard models Points-based weight-of-evidence scorecard, EBM, GLM, RuleFit and monotonic boosters, as templates Weight-of-evidence transformers, monotonic constraints, GLM, RuleFit, decision tree GA2M rating tables, Eureqa GAM, RuleFit, monotonic constraints Linear models, decision tree, monotonic constraints on CatBoost; credit scoring as a packaged solution
Identifiers Card numbers, IBANs, phone numbers, emails, names and IP addresses pseudonymised before any model sees them Not found in public docs Personal-data detection that flags columns to leave out (preview, self-managed installs); pseudonymisation not found in public docs Not found in public docs for tabular models
Release sign-off A release record with the approval, an independent challenge, conditions and limitations Not found in public docs Deployment approval policies an administrator sets up, with a log of policy events Sign-off in Govern, a separate node not in every licence
Clustering Seven: k-means, mini-batch k-means, DBSCAN, HDBSCAN, Gaussian mixture, agglomerative, BIRCH. Cluster count chosen for you; clusters named and explained K-Means (experimental) K-Means, Gaussian mixture, HDBSCAN; separate licence Eight clustering algorithms
Models 15 model variants: LightGBM (5 modes), CatBoost (3), XGBoost (2), random forest, extra trees, linear and two neural nets. Plus glass-box EBM, GLM, RuleFit and scorecard, and the TabICL foundation model. Stacked or deeply tuned ensembles XGBoost, LightGBM, CatBoost, random forest, GLM, decision tree, RuleFit, FTRL, zero-inflated, TensorFlow and PyTorch models (including BERT and GrowNet) GBMs, GLM, GAM rating tables, Keras, Eureqa, RuleFit, SVM, KNN Linear, forests, LightGBM, XGBoost, CatBoost, TabICL, neural nets, SVM, KNN
Classification and regression Yes, with calibration, thresholds, business value and class-imbalance treatment chosen by search Yes Yes Yes
Regression prediction intervals On by default: conformal intervals, by quantile regression where it fits, so the width adapts per row, at 80% and 95%, flagged when the coverage can't be trusted Empirical intervals from holdout predictions Built in for time series; for other regression, a conformal method in a code accelerator A helper model fitted on the residuals, 95% by default
Production Python scoring package that refuses to load in an incompatible version, REST scoring, MLflow including Unity Catalog, Treelite Python, Java and C++ scoring Deployment, monitoring, approvals Deployment, monitoring, Govern
Images No Yes Yes Yes
How it compares with open-source AutoML libraries

These are libraries you code against. ParleML is a platform built on the strongest of them, AutoGluon, so it isn't on this list. As of 18 October 2026, from each project's documentation, release pages and PyPI:

ParleML H2O-3 AutoML mljar-supervised PyCaret FLAML
Status Free platform, closed source Apache-2.0; only 3.46.0.x releases since March 2024 MIT; actively released MIT; last stable release April 2024, 4.0 in alpha MIT; actively released (Microsoft)
How you use it Web app, REST API, agents and MCP; the engines as Python libraries Python, R, or the Flow notebook Python; MLJAR Studio, a desktop app with free and paid plans, adds an AI Data Analyst Python (a dashboard is coming in 4.0, now in alpha) Python
Problems Everything in the other columns, plus ranking, recommendation and basket rules Classification, regression; clustering and Isolation Forest outside AutoML Classification, regression Classification, regression, forecasting, clustering, anomalies Classification, regression, forecasting, ranking, text tasks
Interpretation The full standard set, plus archetypes, tested rules, error tree, ALE, trust summary SHAP on tree models, PDP, ICE, fairness, one-call report SHAP, permutation importance, fairness, Markdown reports SHAP plots, PDP, fairness check Feature importance from the underlying model; no SHAP or explanation tools found
Feature engineering See the feature-engineering section Target encoding Golden features, feature selection Preprocessing pipeline Minimal
Agents Built in, with your own subscription, key or local model None found in public docs Studio only; subscription plans have monthly prompt quotas None found in public docs None found in public docs
Scoring Python package, REST, MLflow, Treelite Java MOJO and POJO Saved Python models In 3.3: cloud and Docker deploy, transpile to Java, C or Go (removed in 4.0) Python objects

Each is good at something: H2O-3 scores in Java with no Python and scales on a cluster; mljar-supervised is clean, actively released and writes clear reports; PyCaret 3.3 deploys in one call; FLAML tunes fast and cheaply, and fine-tunes transformer models for text.

Agents: a different direction from DataRobot and Dataiku

As we read their documentation, DataRobot focuses on a platform for building, hosting and governing your company's own agents (an LLM gateway, guardrails, evaluation, tracing), with open-source skills and an MCP server so coding agents can call it from outside.

Dataiku has its own building agent, Cobuild, run on Dataiku AI Services or your own models, and an agent hub for business agents. Dataiku Headless lets Claude Code, Codex and Cursor drive Cobuild from outside, coding CLIs run in its built-in IDE, and its MCP server exposes the agents you build there.

ParleML takes another road. It doesn't host your agents or sell its own. It makes the agents you already use good at machine learning:

  • The coding agent is the data scientist. Codex, Claude Code or Gemini CLI run whole ML sessions in ParleML: framing, data, experiments, challenges, release. You watch them on a live map and answer their decisions, and every step is on record.
  • Your access, no platform fee. Use your ChatGPT, Claude or Google AI plan, an API key or a local model; ParleML adds no charge on top.
  • Or bring ParleML to your agent. The MCP server exposes the ML platform itself (data, experiments, evaluation, interpretation) to Claude, ChatGPT, Cursor and VS Code, with interactive ParleML panels inside the chat.
  • Free, including every agent feature.
Where the big platforms are still stronger
  • Production deployment and monitoring, with approval workflows. Export to your MLOps tooling instead.
  • Hosting and governing your company's own business agents (DataRobot, Dataiku).
  • Images, and in Dataiku causal and uplift modelling, and visual data preparation.
  • Very large data on a cluster. ParleML runs in one container, from a laptop to a large server, for data that fits one machine.
  • Cloud warehouse connectors. Files and PostgreSQL, MySQL, SQL Server, Oracle and SQLite today.
  • Enterprise support contracts.
Sources

Questions

Does ParleML send anything home?

No. No telemetry, no usage reporting, no update checks. ParleML also switches off the telemetry of the libraries and agent CLIs it runs on.

The only traffic that leaves your machine is what you choose: your agents' calls to the AI provider you signed them in to, a model download the first time you use a model the image doesn't include, and connections you add yourself. Use a local model for the agents and ParleML runs fully offline. How to set it up · Privacy.

Is it free? Is it open source? What do the agents cost?

Free: yes, including commercial use. Exported scoring packages may be sold and distributed with the required ParleML runtime unchanged. Selling ParleML or charging for access to its platform is not permitted. See the license.

Open source: not at the moment. We are considering it.

Agents: ParleML adds nothing on top. Sign the agents in with the ChatGPT, Claude or Google AI plan you already have and they run on it, or use an API key, or a free local model through Ollama (or LM Studio, with Codex). Everything outside agent sessions needs no AI at all.

Next